Sentence feature extraction for machine-generated multiple sequence dialogues

Abstract

Recent advances in deep learning has been leading the development of various methods in higher level technology. Natural Language Processing is one of the areas that has made noticeable achievements through deep learning, especially in machine translation and chatbot algorithms. According to literature, many of the current technologies take advantage of the LSTM architecture to take words sequence by sequence in order to generate sentences. However, many of the current state-of-the-art models have the problem of not being able to process dialogues of more than 2 sequences due to limitations in the architecture. In this paper, we propose an end-to-end dialogue system model that is capable of processing up to 3 sequences of dialogues. We use this dialogue system to extract raw sentence representations, and engineer the raw features with an unsupervised learning method. As a result, we’ve shown that using sentence-level features help increase the overall performance of a non-goal-oriented dialogue system.